Wavelet-based Machine Learning Techniques for ECG Signal Analysis

Abstract

Machine learning of ECG is a core component in any of the ECG-based healthcare informatics system. Since the ECG is a nonlinear signal, the subtle changes in its amplitude and duration are not well manifested in time and frequency domains. Therefore, in this chapter, we introduce a machine-learning approach to screen arrhythmia from normal sinus rhythm from the ECG. The methodology consists of R-point detection using the Pan-Tompkins algorithm, discrete wavelet transform (DWT) decomposition, sub-band principal component analysis (PCA), statistical validation of features, and subsequent pattern classification. The k-fold cross validation is used in order to reduce the bias in choosing training and testing sets for classification. The average accuracy of classification is used as a benchmark for comparison. Different classifiers used are Gaussian mixture model (GMM), error back propagation neural network (EBPNN), and support vector machine (SVM). The DWT basis functions used are Daubechies-4, Daubechies-6, Daubechies-8, Symlet-2, Symlet-4, Symlet-6, Symlet-8, Coiflet-2, and Coiflet-5. An attempt is made to exploit the energy compaction in the wavelet sub-bands to yield higher classification accuracy. Results indicate that the Symlet-2 wavelet basis function provides the highest accuracy in classification. Among the classifiers, SVM yields the highest classification accuracy, whereas EBPNN yields a higher accuracy than GMM. The use of other time frequency representations using different time frequency kernels as a future direction is also observed. The developed machine-learning approach can be used in a web-based telemedicine system, which can be used in remote monitoring of patients in many healthcare informatics systems.